Skip to content

Benchmarks

The best speech emotion recognition model, measured

We built speech-emotion-bench to settle a simple question: which model actually understands emotion in speech? It scores 64,384 held-out clips across roughly 20 languages and 7 emotion classes with one identical pipeline for 27+ systems — open models, frontier APIs, and Oruk Spectra alike. On this set, Oruk Spectra leads at 77.6% accuracy and 0.810 macro F1.

Evaluations · speech-emotion-bench

Best-in-class at understanding people

We built speech-emotion-bench to settle it: 64,384 held-out clips across ~20 languages and seven emotions, one scoring pipeline for all 27+ systems. Oruk Spectra against the best open models and frontier APIs, each given best-chance prompting. It isn't close.

0.0%+8.9

Oruk Spectra · accuracy, 64,384-clip held-out set

Oruk Spectraours · 640M77.6
emotion2vec+ seedbest open68.7
emotion2vec finetunedopen63.6
SenseVoice-smallopen55.7
Gemini 3 Flash PreviewAPI46.0
Gemini 2.5 FlashAPI45.4
GPT-Audio 1.5API43.3
Claude Opus 4.8text only40.3

* Closed models are scored on a fixed 5,000-clip stratified subsample; rescoring open models on the same subsample shifts results by less than two points. Claude receives transcripts only — it has no audio input. Oruk Spectra is trained on the benchmark's training split; every other system is zero-shot cross-corpus.

Per-emotion F1

Macro F1 for each of the seven emotions, Oruk Spectra against the best frontier API (Gemini 3 Flash Preview). The gap is widest on the emotions text-only and general-audio models miss most.

EmotionOruk SpectraBest frontier API
Disgust0.9050.185
Fear0.8640.355
Surprise0.8580.268
Anger0.8060.462
Neutral0.7560.529
Sadness0.7460.313
Happiness0.7370.502

How we score it

A held-out speech emotion recognition benchmark: 64,384 clips across ~20 languages and 7 emotion classes, scored with one identical pipeline for 27+ systems. Closed/API models are scored on a fixed 5,000-clip stratified subsample; rescoring open models on the same subsample shifts results by less than two points.

Closed models receive best-chance prompting: expert-annotator framing, acoustic definitions for all seven classes, temperature 0, and constrained single-label output. Claude models receive transcripts only — they have no audio input.

One fairness note: Oruk Spectra is trained on the benchmark's training split, while every other system is evaluated zero-shot cross-corpus. The harness and results are public at github.com/Oruk-AI/speech-emotion-bench.

Make your agents actually understand people

Join the labs, startups, and enterprises teaching their products to hear what people really mean. Request access to the Oruk API.

© 2026 Oruk Labs. All rights reserved.